src/video/v9990/V9990BitmapConverter.cc at main · doomsdayone/openmsx

◆ main openmsx / src / video / v9990 / V9990BitmapConverter.cc

17.8 KiB · C++ · last changed chore(): initial commit 1 month ago source blame history raw

#include "V9990BitmapConverter.hh"

#include "V9990.hh"
#include "V9990VRAM.hh"

#include "narrow.hh"
#include "unreachable.hh"

#include <array>
#include <cassert>
#include <concepts>
#include <cstdint>

namespace openmsx {

V9990BitmapConverter::V9990BitmapConverter(
		V9990& vdp_,
		std::span<const Pixel,    64> palette64_,  std::span<const int16_t,  64> palette64_32768_,
		std::span<const Pixel,   256> palette256_, std::span<const int16_t, 256> palette256_32768_,
		std::span<const Pixel, 32768> palette32768_)
	: vdp(vdp_), vram(vdp.getVRAM())
	, palette64 (palette64_ ), palette64_32768 (palette64_32768_ )
	, palette256(palette256_), palette256_32768(palette256_32768_)
	, palette32768(palette32768_)
{
	setColorMode(V9990ColorMode::PP, V9990DisplayMode::B0); // initialize with dummy values
}

template<bool YJK, bool PAL, bool SKIP, std::unsigned_integral Pixel, typename ColorLookup>
static inline void draw_YJK_YUV_PAL(
	ColorLookup color, const V9990VRAM& vram,
	Pixel* __restrict& out, unsigned& address, int firstX = 0)
{
	std::array<uint8_t, 4> data;
	for (auto& d : data) {
		d = vram.readVRAMBx(address++);
	}

	int u = (data[2] & 7) + ((data[3] & 3) << 3) - ((data[3] & 4) << 3);
	int v = (data[0] & 7) + ((data[1] & 3) << 3) - ((data[1] & 4) << 3);

	for (auto i : xrange(SKIP ? firstX : 0, 4)) {
		if (PAL && (data[i] & 0x08)) {
			*out++ = color.lookup64(data[i] >> 4);
		} else {
			int y = (data[i] & 0xF8) >> 3;
			int r = std::clamp(y + u,                   0, 31);
			int g = std::clamp((5 * y - 2 * u - v) / 4, 0, 31);
			int b = std::clamp(y + v,                   0, 31);
			// The only difference between YUV and YJK is that
			// green and blue are swapped.
			if constexpr (YJK) std::swap(g, b);
			*out++ = color.lookup32768((g << 10) + (r << 5) + b);
		}
	}
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBYUV(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = (x & ~3) + y * vdp.getImageWidth();
	if (x & 3) {
		draw_YJK_YUV_PAL<false, false, true>(
			color, vram, out, address, x & 3);
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		draw_YJK_YUV_PAL<false, false, false>(
			color, vram, out, address);
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBYUVP(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	// TODO this mode cannot be shown in B4 and higher resolution modes
	//      (So the dual palette for B4 modes is not an issue here.)
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = (x & ~3) + y * vdp.getImageWidth();
	if (x & 3) {
		draw_YJK_YUV_PAL<false, true, true>(
			color, vram, out, address, x & 3);
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		draw_YJK_YUV_PAL<false, true, false>(
			color, vram, out, address);
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBYJK(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = (x & ~3) + y * vdp.getImageWidth();
	if (x & 3) {
		draw_YJK_YUV_PAL<true, false, true>(
			color, vram, out, address, x & 3);
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		draw_YJK_YUV_PAL<true, false, false>(
			color, vram, out, address);
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBYJKP(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	// TODO this mode cannot be shown in B4 and higher resolution modes
	//      (So the dual palette for B4 modes is not an issue here.)
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = (x & ~3) + y * vdp.getImageWidth();
	if (x & 3) {
		draw_YJK_YUV_PAL<true, true, true>(
			color, vram, out, address, x & 3);
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		draw_YJK_YUV_PAL<true, true, false>(
			color, vram, out, address);
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBD16(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = 2 * (x + y * vdp.getImageWidth());
	if (vdp.isSuperimposing()) {
		auto transparent = color.lookup256(0);
		for (/**/; nrPixels > 0; --nrPixels) {
			uint8_t high = vram.readVRAMBx(address + 1);
			if (high & 0x80) {
				*out = transparent;
			} else {
				uint8_t low  = vram.readVRAMBx(address + 0);
				*out = color.lookup32768(low + 256 * high);
			}
			address += 2;
			out += 1;
		}
	} else {
		for (/**/; nrPixels > 0; --nrPixels) {
			uint8_t low  = vram.readVRAMBx(address++);
			uint8_t high = vram.readVRAMBx(address++);
			*out++ = color.lookup32768((low + 256 * high) & 0x7FFF);
		}
	}
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBD8(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = x + y * vdp.getImageWidth();
	for (/**/; nrPixels > 0; --nrPixels) {
		*out++ = color.lookup256(vram.readVRAMBx(address++));
	}
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBP6(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = x + y * vdp.getImageWidth();
	for (/**/; nrPixels > 0; --nrPixels) {
		*out++ = color.lookup64(vram.readVRAMBx(address++) & 0x3F);
	}
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBP4(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	assert(nrPixels > 0);
	unsigned address = (x + y * vdp.getImageWidth()) / 2;
	color.set64Offset((vdp.getPaletteOffset() & 0xC) << 2);
	if (x & 1) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64(data & 0x0F);
		--nrPixels;
	}
	for (/**/; nrPixels > 0; nrPixels -= 2) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64(data >> 4);
		*out++ = color.lookup64(data & 0x0F);
	}
	// Note: this possibly draws 1 pixel too many, but that's ok.
}
template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBP4HiRes(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	// Verified on real HW:
	//   Bit PLT05 in palette offset is ignored, instead for even pixels
	//   bit 'PLT05' is '0', for odd pixels it's '1'.
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	unsigned address = (x + y * vdp.getImageWidth()) / 2;
	color.set64Offset((vdp.getPaletteOffset() & 0x4) << 2);
	if (x & 1) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64(32 | (data & 0x0F));
		--nrPixels;
	}
	for (/**/; nrPixels > 0; nrPixels -= 2) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64( 0 | (data >> 4  ));
		*out++ = color.lookup64(32 | (data & 0x0F));
	}
	// Note: this possibly draws 1 pixel too many, but that's ok.
}

template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBP2(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	assert(nrPixels > 0);
	unsigned address = (x + y * vdp.getImageWidth()) / 4;
	color.set64Offset(vdp.getPaletteOffset() << 2);
	if (x & 3) {
		uint8_t data = vram.readVRAMBx(address++);
		if ((x & 3) <= 1) *out++ = color.lookup64((data & 0x30) >> 4);
		if ((x & 3) <= 2) *out++ = color.lookup64((data & 0x0C) >> 2);
		if (true)         *out++ = color.lookup64((data & 0x03) >> 0);
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64((data & 0xC0) >> 6);
		*out++ = color.lookup64((data & 0x30) >> 4);
		*out++ = color.lookup64((data & 0x0C) >> 2);
		*out++ = color.lookup64((data & 0x03) >> 0);
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}
template<std::unsigned_integral Pixel, typename ColorLookup>
static void rasterBP2HiRes(
	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	std::span<Pixel> buf, unsigned x, unsigned y)
{
	// Verified on real HW:
	//   Bit PLT05 in palette offset is ignored, instead for even pixels
	//   bit 'PLT05' is '0', for odd pixels it's '1'.
	Pixel* __restrict out = buf.data();
	int nrPixels = narrow<int>(buf.size());
	assert(nrPixels > 0);
	unsigned address = (x + y * vdp.getImageWidth()) / 4;
	color.set64Offset((vdp.getPaletteOffset() & 0x7) << 2);
	if (x & 3) {
		uint8_t data = vram.readVRAMBx(address++);
		if ((x & 3) <= 1) *out++ = color.lookup64(32 | ((data & 0x30) >> 4));
		if ((x & 3) <= 2) *out++ = color.lookup64( 0 | ((data & 0x0C) >> 2));
		if (true)         *out++ = color.lookup64(32 | ((data & 0x03) >> 0));
		nrPixels -= narrow<int>(4 - (x & 3));
	}
	for (/**/; nrPixels > 0; nrPixels -= 4) {
		uint8_t data = vram.readVRAMBx(address++);
		*out++ = color.lookup64( 0 | ((data & 0xC0) >> 6));
		*out++ = color.lookup64(32 | ((data & 0x30) >> 4));
		*out++ = color.lookup64( 0 | ((data & 0x0C) >> 2));
		*out++ = color.lookup64(32 | ((data & 0x03) >> 0));
	}
	// Note: this can draw up to 3 pixels too many, but that's ok.
}

// Helper class to translate V9990 palette indices into host Pixel values.
template<std::unsigned_integral Pixel>
class PaletteLookup
{
public:
	PaletteLookup(std::span<const Pixel,    64> palette64_,
	              std::span<const Pixel,   256> palette256_,
	              std::span<const Pixel, 32768> palette32768_)
	        : palette64Base(palette64_)
	        , palette64(palette64_)
	        , palette256(palette256_)
	        , palette32768(palette32768_)
	{
	}

	void set64Offset(size_t offset) { palette64 = palette64Base.subspan(offset); }
	[[nodiscard]] Pixel lookup64   (size_t idx) const { return palette64   [idx]; }
	[[nodiscard]] Pixel lookup256  (size_t idx) const { return palette256  [idx]; }
	[[nodiscard]] Pixel lookup32768(size_t idx) const { return palette32768[idx]; }

private:
	std::span<const Pixel,    64> palette64Base;
	std::span<const Pixel>        palette64;
	std::span<const Pixel,   256> palette256;
	std::span<const Pixel, 32768> palette32768;
};

// Helper class to translate V9990 palette indices (64-entry, 256-entry and
// 32768-entry palettes) into V9990 32768-entry palette indices.
class IndexLookup
{
public:
	IndexLookup(std::span<const int16_t, 64> palette64_, std::span<const int16_t, 256> palette256_)
	        : palette64_32768Base(palette64_)
	        , palette64_32768(palette64_)
	        , palette256_32768(palette256_)
	{
	}

	void set64Offset(size_t offset) { palette64_32768 = palette64_32768Base.subspan(offset); }
	[[nodiscard]] int16_t lookup64   (size_t idx) const { return palette64_32768 [idx]; }
	[[nodiscard]] int16_t lookup256  (size_t idx) const { return palette256_32768[idx]; }
	[[nodiscard]] int16_t lookup32768(size_t idx) const { return int16_t(idx); }

private:
	std::span<const int16_t,  64> palette64_32768Base;
	std::span<const int16_t>      palette64_32768;
	std::span<const int16_t, 256> palette256_32768;
};

template<std::unsigned_integral Pixel, typename ColorLookup>
static void raster(V9990ColorMode colorMode, bool highRes,
                   ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
                   std::span<Pixel> out, unsigned x, unsigned y)
{
	switch (colorMode) {
	using enum V9990ColorMode;
	case BYUV:  return rasterBYUV <Pixel>(color, vdp, vram, out, x, y);
	case BYUVP: return rasterBYUVP<Pixel>(color, vdp, vram, out, x, y);
	case BYJK:  return rasterBYJK <Pixel>(color, vdp, vram, out, x, y);
	case BYJKP: return rasterBYJKP<Pixel>(color, vdp, vram, out, x, y);
	case BD16:  return rasterBD16 <Pixel>(color, vdp, vram, out, x, y);
	case BD8:   return rasterBD8  <Pixel>(color, vdp, vram, out, x, y);
	case BP6:   return rasterBP6  <Pixel>(color, vdp, vram, out, x, y);
	case BP4:   return highRes ? rasterBP4HiRes<Pixel>(color, vdp, vram, out, x, y)
	                           : rasterBP4     <Pixel>(color, vdp, vram, out, x, y);
	case BP2:   return highRes ? rasterBP2HiRes<Pixel>(color, vdp, vram, out, x, y)
	                           : rasterBP2     <Pixel>(color, vdp, vram, out, x, y);
	default:    UNREACHABLE;
	}
}

// Missing details in the V9990 application manual (reverse engineered from
// tests on a real gfx9000):
//  * Cursor 0 is drawn on top of cursor 1  (IOW cursor 0 has higher priority).
//    This remains the case when both cursors use the 'EOR' feature (it's not
//    so that EOR is applied twice).
//  * The CC1,CC0,EOR bits in the cursor attribute table work like this:
//     (CC1,CC0):
//       when (0,0): pick the resulting color from bitmap rendering (this is
//                   a 15 bit RGB value)
//       when (x,y): pick the color from palette with index R#28:x:y (this is
//                   also a 15 bit RGB color)
//     (EOR):
//       when 0: use the above color unchanged
//       when 1: flip all the bits in the above color (IOW XOR with 0x7fff)
//    From this follows:
//      (CC1,CC0,EOR)==(0,0,0):
//        Results in an invisible cursor: each pixel is colored the same as the
//        corresponding background pixel.
//      (CC1,CC0,EOR)==(0,0,1):
//        This is the only combination where the cursor is drawn using multiple
//        colors, each pixel is the complement of the corresponding background
//        pixel.
//      (CC1,CC0,EOR)==(x,y,0):
//        This is the 'usual' configuration, cursor is drawn with a specific
//        color from the palette (also when bitmap rendering is not using the
//        palette, e.g. YJK or BD8 mode).
//      (CC1,CC0,EOR)==(x,y,1):
//        This undocumented mode draws the cursor with a single color which is
//        the complement of a specific palette color.
class CursorInfo
{
public:
	CursorInfo(const V9990& vdp, const V9990VRAM& vram, std::span<const int16_t, 64> palette64_32768,
	           unsigned attrAddr, unsigned patAddr,
		   int displayY, bool drawCursor)
	{
		if (!drawCursor) return;

		unsigned attrY = vram.readVRAMBx(attrAddr + 0) +
		                (vram.readVRAMBx(attrAddr + 2) & 1) * 256;
		attrY += vdp.isInterlaced() ? 2 : 1; // 1 or 2 lines later
		unsigned cursorLine = (displayY - attrY) & 511;
		if (cursorLine >= 32) return;

		uint8_t attr = vram.readVRAMBx(attrAddr + 6);
		if ((attr & 0x10) || ((attr & 0xe0) == 0x00)) {
			// don't display
			return;
		}

		pattern = (vram.readVRAMBx(patAddr + 4 * cursorLine + 0) << 24)
		        + (vram.readVRAMBx(patAddr + 4 * cursorLine + 1) << 16)
		        + (vram.readVRAMBx(patAddr + 4 * cursorLine + 2) <<  8)
		        + (vram.readVRAMBx(patAddr + 4 * cursorLine + 3) <<  0);
		if (pattern == 0) {
			// optimization, completely transparent line
			return;
		}

		// mark cursor visible
		x = vram.readVRAMBx(attrAddr + 4) + (attr & 3) * 256;

		doXor = (attr & 0xe0) == 0x20;

		auto colorIdx = vdp.getSpritePaletteOffset() + (attr >> 6);
		color = palette64_32768[colorIdx];
		if (attr & 0x20) color ^= 0x7fff;
	}

	[[nodiscard]] bool isVisible() const {
		return x != unsigned(-1);
	}
	[[nodiscard]] bool dot() const {
		return (x == 0) && (pattern & 0x80000000);
	}
	void shift() {
		if (x) {
			--x;
		} else {
			pattern <<= 1;
		}
	}

public:
	unsigned x{unsigned(-1)}; // not visible
	uint32_t pattern{0};
	int16_t color{0};
	bool doXor{false};
};

void V9990BitmapConverter::convertLine(
	std::span<Pixel> dst, unsigned x, unsigned y,
	int cursorY, bool drawCursors) const
{
	assert(dst.size() <= 1024);

	CursorInfo cursor0(vdp, vram, palette64_32768, 0x7fe00, 0x7ff00, cursorY, drawCursors);
	CursorInfo cursor1(vdp, vram, palette64_32768, 0x7fe08, 0x7ff80, cursorY, drawCursors);

	if (cursor0.isVisible() || cursor1.isVisible()) {
		// raster background into a temporary buffer
		std::array<uint16_t, 1024 + 3> buf; // allow to draw upto 3 pixels too many, e.g. see rasterBP2()
		raster(colorMode, highRes,
		       IndexLookup(palette64_32768, palette256_32768),
		       vdp, vram,
		       subspan(buf, 0, dst.size()), x, y);

		// draw sprites in this buffer
		// TODO can be optimized
		// TODO probably goes wrong when startX != 0
		// TODO investigate dual palette in B4 and higher modes
		// TODO check X-roll behavior
		for (auto i : xrange(dst.size())) {
			if (cursor0.dot()) {
				if (cursor0.doXor) {
					buf[i] ^= 0x7fff;
				} else {
					buf[i] = cursor0.color;
				}
			} else if (cursor1.dot()) {
				if (cursor1.doXor) {
					buf[i] ^= 0x7fff;
				} else {
					buf[i] = cursor1.color;
				}
			}
			cursor0.shift();
			cursor1.shift();
			if ((cursor0.pattern == 0) && (cursor1.pattern == 0)) break;
		}

		// copy buffer to destination, translate from V9990 to host colors
		for (auto i : xrange(dst.size())) {
			dst[i] = palette32768[buf[i]];
		}
	} else {
		// Optimization: no cursor(s) visible on this line, directly draw to destination
		raster(colorMode, highRes,
		       PaletteLookup<Pixel>(palette64, palette256, palette32768),
		       vdp, vram, dst, x, y);
	}
}

} // namespace openmsx

`/`	Search
`g` `c`	Code
`g` `h`	History
`g` `b`	Bookmarks
`g` `t`	Trust
`y`	Permalink to this state (files and trees)
`?`	This sheet

e8555c89 Andy McCall 1 month ago	1	#include "V9990BitmapConverter.hh"
	2
	3	#include "V9990.hh"
	4	#include "V9990VRAM.hh"
	5
	6	#include "narrow.hh"
	7	#include "unreachable.hh"
	8
	9	#include <array>
	10	#include <cassert>
	11	#include <concepts>
	12	#include <cstdint>
	13
	14	namespace openmsx {
	15
	16	V9990BitmapConverter::V9990BitmapConverter(
	17	V9990& vdp_,
	18	std::span<const Pixel, 64> palette64_, std::span<const int16_t, 64> palette64_32768_,
	19	std::span<const Pixel, 256> palette256_, std::span<const int16_t, 256> palette256_32768_,
	20	std::span<const Pixel, 32768> palette32768_)
	21	: vdp(vdp_), vram(vdp.getVRAM())
	22	, palette64 (palette64_ ), palette64_32768 (palette64_32768_ )
	23	, palette256(palette256_), palette256_32768(palette256_32768_)
	24	, palette32768(palette32768_)
	25	{
	26	setColorMode(V9990ColorMode::PP, V9990DisplayMode::B0); // initialize with dummy values
	27	}
	28
	29	template<bool YJK, bool PAL, bool SKIP, std::unsigned_integral Pixel, typename ColorLookup>
	30	static inline void draw_YJK_YUV_PAL(
	31	ColorLookup color, const V9990VRAM& vram,
	32	Pixel* __restrict& out, unsigned& address, int firstX = 0)
	33	{
	34	std::array<uint8_t, 4> data;
	35	for (auto& d : data) {
	36	d = vram.readVRAMBx(address++);
	37	}
	38
	39	int u = (data[2] & 7) + ((data[3] & 3) << 3) - ((data[3] & 4) << 3);
	40	int v = (data[0] & 7) + ((data[1] & 3) << 3) - ((data[1] & 4) << 3);
	41
	42	for (auto i : xrange(SKIP ? firstX : 0, 4)) {
	43	if (PAL && (data[i] & 0x08)) {
	44	*out++ = color.lookup64(data[i] >> 4);
	45	} else {
	46	int y = (data[i] & 0xF8) >> 3;
	47	int r = std::clamp(y + u, 0, 31);
	48	int g = std::clamp((5 * y - 2 * u - v) / 4, 0, 31);
	49	int b = std::clamp(y + v, 0, 31);
	50	// The only difference between YUV and YJK is that
	51	// green and blue are swapped.
	52	if constexpr (YJK) std::swap(g, b);
	53	*out++ = color.lookup32768((g << 10) + (r << 5) + b);
	54	}
	55	}
	56	}
	57
	58	template<std::unsigned_integral Pixel, typename ColorLookup>
	59	static void rasterBYUV(
	60	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
	61	std::span<Pixel> buf, unsigned x, unsigned y)
	62	{
	63	Pixel* __restrict out = buf.data();
	64	int nrPixels = narrow<int>(buf.size());
65	unsigned address = (x & ~3) + y * vdp.getImageWidth();
66	if (x & 3) {
67	draw_YJK_YUV_PAL<false, false, true>(
68	color, vram, out, address, x & 3);
69	nrPixels -= narrow<int>(4 - (x & 3));
70	}
71	for (/**/; nrPixels > 0; nrPixels -= 4) {
72	draw_YJK_YUV_PAL<false, false, false>(
73	color, vram, out, address);
74	}
75	// Note: this can draw up to 3 pixels too many, but that's ok.
76	}
77
78	template<std::unsigned_integral Pixel, typename ColorLookup>
79	static void rasterBYUVP(
80	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
81	std::span<Pixel> buf, unsigned x, unsigned y)
82	{
83	// TODO this mode cannot be shown in B4 and higher resolution modes
84	// (So the dual palette for B4 modes is not an issue here.)
85	Pixel* __restrict out = buf.data();
86	int nrPixels = narrow<int>(buf.size());
87	unsigned address = (x & ~3) + y * vdp.getImageWidth();
88	if (x & 3) {
89	draw_YJK_YUV_PAL<false, true, true>(
90	color, vram, out, address, x & 3);
91	nrPixels -= narrow<int>(4 - (x & 3));
92	}
93	for (/**/; nrPixels > 0; nrPixels -= 4) {
94	draw_YJK_YUV_PAL<false, true, false>(
95	color, vram, out, address);
96	}
97	// Note: this can draw up to 3 pixels too many, but that's ok.
98	}
99
100	template<std::unsigned_integral Pixel, typename ColorLookup>
101	static void rasterBYJK(
102	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
103	std::span<Pixel> buf, unsigned x, unsigned y)
104	{
105	Pixel* __restrict out = buf.data();
106	int nrPixels = narrow<int>(buf.size());
107	unsigned address = (x & ~3) + y * vdp.getImageWidth();
108	if (x & 3) {
109	draw_YJK_YUV_PAL<true, false, true>(
110	color, vram, out, address, x & 3);
111	nrPixels -= narrow<int>(4 - (x & 3));
112	}
113	for (/**/; nrPixels > 0; nrPixels -= 4) {
114	draw_YJK_YUV_PAL<true, false, false>(
115	color, vram, out, address);
116	}
117	// Note: this can draw up to 3 pixels too many, but that's ok.
118	}
119
120	template<std::unsigned_integral Pixel, typename ColorLookup>
121	static void rasterBYJKP(
122	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
123	std::span<Pixel> buf, unsigned x, unsigned y)
124	{
125	// TODO this mode cannot be shown in B4 and higher resolution modes
126	// (So the dual palette for B4 modes is not an issue here.)
127	Pixel* __restrict out = buf.data();
128	int nrPixels = narrow<int>(buf.size());
129	unsigned address = (x & ~3) + y * vdp.getImageWidth();
130	if (x & 3) {
131	draw_YJK_YUV_PAL<true, true, true>(
132	color, vram, out, address, x & 3);
133	nrPixels -= narrow<int>(4 - (x & 3));
134	}
135	for (/**/; nrPixels > 0; nrPixels -= 4) {
136	draw_YJK_YUV_PAL<true, true, false>(
137	color, vram, out, address);
138	}
139	// Note: this can draw up to 3 pixels too many, but that's ok.
140	}
141
142	template<std::unsigned_integral Pixel, typename ColorLookup>
143	static void rasterBD16(
144	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
145	std::span<Pixel> buf, unsigned x, unsigned y)
146	{
147	Pixel* __restrict out = buf.data();
148	int nrPixels = narrow<int>(buf.size());
149	unsigned address = 2 * (x + y * vdp.getImageWidth());
150	if (vdp.isSuperimposing()) {
151	auto transparent = color.lookup256(0);
152	for (/**/; nrPixels > 0; --nrPixels) {
153	uint8_t high = vram.readVRAMBx(address + 1);
154	if (high & 0x80) {
155	*out = transparent;
156	} else {
157	uint8_t low = vram.readVRAMBx(address + 0);
158	out = color.lookup32768(low + 256 high);
159	}
160	address += 2;
161	out += 1;
162	}
163	} else {
164	for (/**/; nrPixels > 0; --nrPixels) {
165	uint8_t low = vram.readVRAMBx(address++);
166	uint8_t high = vram.readVRAMBx(address++);
167	out++ = color.lookup32768((low + 256 high) & 0x7FFF);
168	}
169	}
170	}
171
172	template<std::unsigned_integral Pixel, typename ColorLookup>
173	static void rasterBD8(
174	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
175	std::span<Pixel> buf, unsigned x, unsigned y)
176	{
177	Pixel* __restrict out = buf.data();
178	int nrPixels = narrow<int>(buf.size());
179	unsigned address = x + y * vdp.getImageWidth();
180	for (/**/; nrPixels > 0; --nrPixels) {
181	*out++ = color.lookup256(vram.readVRAMBx(address++));
182	}
183	}
184
185	template<std::unsigned_integral Pixel, typename ColorLookup>
186	static void rasterBP6(
187	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
188	std::span<Pixel> buf, unsigned x, unsigned y)
189	{
190	Pixel* __restrict out = buf.data();
191	int nrPixels = narrow<int>(buf.size());
192	unsigned address = x + y * vdp.getImageWidth();
193	for (/**/; nrPixels > 0; --nrPixels) {
194	*out++ = color.lookup64(vram.readVRAMBx(address++) & 0x3F);
195	}
196	}
197
198	template<std::unsigned_integral Pixel, typename ColorLookup>
199	static void rasterBP4(
200	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
201	std::span<Pixel> buf, unsigned x, unsigned y)
202	{
203	Pixel* __restrict out = buf.data();
204	int nrPixels = narrow<int>(buf.size());
205	assert(nrPixels > 0);
206	unsigned address = (x + y * vdp.getImageWidth()) / 2;
207	color.set64Offset((vdp.getPaletteOffset() & 0xC) << 2);
208	if (x & 1) {
209	uint8_t data = vram.readVRAMBx(address++);
210	*out++ = color.lookup64(data & 0x0F);
211	--nrPixels;
212	}
213	for (/**/; nrPixels > 0; nrPixels -= 2) {
214	uint8_t data = vram.readVRAMBx(address++);
215	*out++ = color.lookup64(data >> 4);
216	*out++ = color.lookup64(data & 0x0F);
217	}
218	// Note: this possibly draws 1 pixel too many, but that's ok.
219	}
220	template<std::unsigned_integral Pixel, typename ColorLookup>
221	static void rasterBP4HiRes(
222	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
223	std::span<Pixel> buf, unsigned x, unsigned y)
224	{
225	// Verified on real HW:
226	// Bit PLT05 in palette offset is ignored, instead for even pixels
227	// bit 'PLT05' is '0', for odd pixels it's '1'.
228	Pixel* __restrict out = buf.data();
229	int nrPixels = narrow<int>(buf.size());
230	unsigned address = (x + y * vdp.getImageWidth()) / 2;
231	color.set64Offset((vdp.getPaletteOffset() & 0x4) << 2);
232	if (x & 1) {
233	uint8_t data = vram.readVRAMBx(address++);
234	*out++ = color.lookup64(32 \| (data & 0x0F));
235	--nrPixels;
236	}
237	for (/**/; nrPixels > 0; nrPixels -= 2) {
238	uint8_t data = vram.readVRAMBx(address++);
239	*out++ = color.lookup64( 0 \| (data >> 4 ));
240	*out++ = color.lookup64(32 \| (data & 0x0F));
241	}
242	// Note: this possibly draws 1 pixel too many, but that's ok.
243	}
244
245	template<std::unsigned_integral Pixel, typename ColorLookup>
246	static void rasterBP2(
247	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
248	std::span<Pixel> buf, unsigned x, unsigned y)
249	{
250	Pixel* __restrict out = buf.data();
251	int nrPixels = narrow<int>(buf.size());
252	assert(nrPixels > 0);
253	unsigned address = (x + y * vdp.getImageWidth()) / 4;
254	color.set64Offset(vdp.getPaletteOffset() << 2);
255	if (x & 3) {
256	uint8_t data = vram.readVRAMBx(address++);
257	if ((x & 3) <= 1) *out++ = color.lookup64((data & 0x30) >> 4);
258	if ((x & 3) <= 2) *out++ = color.lookup64((data & 0x0C) >> 2);
259	if (true) *out++ = color.lookup64((data & 0x03) >> 0);
260	nrPixels -= narrow<int>(4 - (x & 3));
261	}
262	for (/**/; nrPixels > 0; nrPixels -= 4) {
263	uint8_t data = vram.readVRAMBx(address++);
264	*out++ = color.lookup64((data & 0xC0) >> 6);
265	*out++ = color.lookup64((data & 0x30) >> 4);
266	*out++ = color.lookup64((data & 0x0C) >> 2);
267	*out++ = color.lookup64((data & 0x03) >> 0);
268	}
269	// Note: this can draw up to 3 pixels too many, but that's ok.
270	}
271	template<std::unsigned_integral Pixel, typename ColorLookup>
272	static void rasterBP2HiRes(
273	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
274	std::span<Pixel> buf, unsigned x, unsigned y)
275	{
276	// Verified on real HW:
277	// Bit PLT05 in palette offset is ignored, instead for even pixels
278	// bit 'PLT05' is '0', for odd pixels it's '1'.
279	Pixel* __restrict out = buf.data();
280	int nrPixels = narrow<int>(buf.size());
281	assert(nrPixels > 0);
282	unsigned address = (x + y * vdp.getImageWidth()) / 4;
283	color.set64Offset((vdp.getPaletteOffset() & 0x7) << 2);
284	if (x & 3) {
285	uint8_t data = vram.readVRAMBx(address++);
286	if ((x & 3) <= 1) *out++ = color.lookup64(32 \| ((data & 0x30) >> 4));
287	if ((x & 3) <= 2) *out++ = color.lookup64( 0 \| ((data & 0x0C) >> 2));
288	if (true) *out++ = color.lookup64(32 \| ((data & 0x03) >> 0));
289	nrPixels -= narrow<int>(4 - (x & 3));
290	}
291	for (/**/; nrPixels > 0; nrPixels -= 4) {
292	uint8_t data = vram.readVRAMBx(address++);
293	*out++ = color.lookup64( 0 \| ((data & 0xC0) >> 6));
294	*out++ = color.lookup64(32 \| ((data & 0x30) >> 4));
295	*out++ = color.lookup64( 0 \| ((data & 0x0C) >> 2));
296	*out++ = color.lookup64(32 \| ((data & 0x03) >> 0));
297	}
298	// Note: this can draw up to 3 pixels too many, but that's ok.
299	}
300
301	// Helper class to translate V9990 palette indices into host Pixel values.
302	template<std::unsigned_integral Pixel>
303	class PaletteLookup
304	{
305	public:
306	PaletteLookup(std::span<const Pixel, 64> palette64_,
307	std::span<const Pixel, 256> palette256_,
308	std::span<const Pixel, 32768> palette32768_)
309	: palette64Base(palette64_)
310	, palette64(palette64_)
311	, palette256(palette256_)
312	, palette32768(palette32768_)
313	{
314	}
315
316	void set64Offset(size_t offset) { palette64 = palette64Base.subspan(offset); }
317	[[nodiscard]] Pixel lookup64 (size_t idx) const { return palette64 [idx]; }
318	[[nodiscard]] Pixel lookup256 (size_t idx) const { return palette256 [idx]; }
319	[[nodiscard]] Pixel lookup32768(size_t idx) const { return palette32768[idx]; }
320
321	private:
322	std::span<const Pixel, 64> palette64Base;
323	std::span<const Pixel> palette64;
324	std::span<const Pixel, 256> palette256;
325	std::span<const Pixel, 32768> palette32768;
326	};
327
328	// Helper class to translate V9990 palette indices (64-entry, 256-entry and
329	// 32768-entry palettes) into V9990 32768-entry palette indices.
330	class IndexLookup
331	{
332	public:
333	IndexLookup(std::span<const int16_t, 64> palette64_, std::span<const int16_t, 256> palette256_)
334	: palette64_32768Base(palette64_)
335	, palette64_32768(palette64_)
336	, palette256_32768(palette256_)
337	{
338	}
339
340	void set64Offset(size_t offset) { palette64_32768 = palette64_32768Base.subspan(offset); }
341	[[nodiscard]] int16_t lookup64 (size_t idx) const { return palette64_32768 [idx]; }
342	[[nodiscard]] int16_t lookup256 (size_t idx) const { return palette256_32768[idx]; }
343	[[nodiscard]] int16_t lookup32768(size_t idx) const { return int16_t(idx); }
344
345	private:
346	std::span<const int16_t, 64> palette64_32768Base;
347	std::span<const int16_t> palette64_32768;
348	std::span<const int16_t, 256> palette256_32768;
349	};
350
351	template<std::unsigned_integral Pixel, typename ColorLookup>
352	static void raster(V9990ColorMode colorMode, bool highRes,
353	ColorLookup color, const V9990& vdp, const V9990VRAM& vram,
354	std::span<Pixel> out, unsigned x, unsigned y)
355	{
356	switch (colorMode) {
357	using enum V9990ColorMode;
358	case BYUV: return rasterBYUV <Pixel>(color, vdp, vram, out, x, y);
359	case BYUVP: return rasterBYUVP<Pixel>(color, vdp, vram, out, x, y);
360	case BYJK: return rasterBYJK <Pixel>(color, vdp, vram, out, x, y);
361	case BYJKP: return rasterBYJKP<Pixel>(color, vdp, vram, out, x, y);
362	case BD16: return rasterBD16 <Pixel>(color, vdp, vram, out, x, y);
363	case BD8: return rasterBD8 <Pixel>(color, vdp, vram, out, x, y);
364	case BP6: return rasterBP6 <Pixel>(color, vdp, vram, out, x, y);
365	case BP4: return highRes ? rasterBP4HiRes<Pixel>(color, vdp, vram, out, x, y)
366	: rasterBP4 <Pixel>(color, vdp, vram, out, x, y);
367	case BP2: return highRes ? rasterBP2HiRes<Pixel>(color, vdp, vram, out, x, y)
368	: rasterBP2 <Pixel>(color, vdp, vram, out, x, y);
369	default: UNREACHABLE;
370	}
371	}
372
373	// Missing details in the V9990 application manual (reverse engineered from
374	// tests on a real gfx9000):
375	// * Cursor 0 is drawn on top of cursor 1 (IOW cursor 0 has higher priority).
376	// This remains the case when both cursors use the 'EOR' feature (it's not
377	// so that EOR is applied twice).
378	// * The CC1,CC0,EOR bits in the cursor attribute table work like this:
379	// (CC1,CC0):
380	// when (0,0): pick the resulting color from bitmap rendering (this is
381	// a 15 bit RGB value)
382	// when (x,y): pick the color from palette with index R#28:x:y (this is
383	// also a 15 bit RGB color)
384	// (EOR):
385	// when 0: use the above color unchanged
386	// when 1: flip all the bits in the above color (IOW XOR with 0x7fff)
387	// From this follows:
388	// (CC1,CC0,EOR)==(0,0,0):
389	// Results in an invisible cursor: each pixel is colored the same as the
390	// corresponding background pixel.
391	// (CC1,CC0,EOR)==(0,0,1):
392	// This is the only combination where the cursor is drawn using multiple
393	// colors, each pixel is the complement of the corresponding background
394	// pixel.
395	// (CC1,CC0,EOR)==(x,y,0):
396	// This is the 'usual' configuration, cursor is drawn with a specific
397	// color from the palette (also when bitmap rendering is not using the
398	// palette, e.g. YJK or BD8 mode).
399	// (CC1,CC0,EOR)==(x,y,1):
400	// This undocumented mode draws the cursor with a single color which is
401	// the complement of a specific palette color.
402	class CursorInfo
403	{
404	public:
405	CursorInfo(const V9990& vdp, const V9990VRAM& vram, std::span<const int16_t, 64> palette64_32768,
406	unsigned attrAddr, unsigned patAddr,
407	int displayY, bool drawCursor)
408	{
409	if (!drawCursor) return;
410
411	unsigned attrY = vram.readVRAMBx(attrAddr + 0) +
412	(vram.readVRAMBx(attrAddr + 2) & 1) * 256;
413	attrY += vdp.isInterlaced() ? 2 : 1; // 1 or 2 lines later
414	unsigned cursorLine = (displayY - attrY) & 511;
415	if (cursorLine >= 32) return;
416
417	uint8_t attr = vram.readVRAMBx(attrAddr + 6);
418	if ((attr & 0x10) \|\| ((attr & 0xe0) == 0x00)) {
419	// don't display
420	return;
421	}
422
423	pattern = (vram.readVRAMBx(patAddr + 4 * cursorLine + 0) << 24)
424	+ (vram.readVRAMBx(patAddr + 4 * cursorLine + 1) << 16)
425	+ (vram.readVRAMBx(patAddr + 4 * cursorLine + 2) << 8)
426	+ (vram.readVRAMBx(patAddr + 4 * cursorLine + 3) << 0);
427	if (pattern == 0) {
428	// optimization, completely transparent line
429	return;
430	}
431
432	// mark cursor visible
433	x = vram.readVRAMBx(attrAddr + 4) + (attr & 3) * 256;
434
435	doXor = (attr & 0xe0) == 0x20;
436
437	auto colorIdx = vdp.getSpritePaletteOffset() + (attr >> 6);
438	color = palette64_32768[colorIdx];
439	if (attr & 0x20) color ^= 0x7fff;
440	}
441
442	[[nodiscard]] bool isVisible() const {
443	return x != unsigned(-1);
444	}
445	[[nodiscard]] bool dot() const {
446	return (x == 0) && (pattern & 0x80000000);
447	}
448	void shift() {
449	if (x) {
450	--x;
451	} else {
452	pattern <<= 1;
453	}
454	}
455
456	public:
457	unsigned x{unsigned(-1)}; // not visible
458	uint32_t pattern{0};
459	int16_t color{0};
460	bool doXor{false};
461	};
462
463	void V9990BitmapConverter::convertLine(
464	std::span<Pixel> dst, unsigned x, unsigned y,
465	int cursorY, bool drawCursors) const
466	{
467	assert(dst.size() <= 1024);
468
469	CursorInfo cursor0(vdp, vram, palette64_32768, 0x7fe00, 0x7ff00, cursorY, drawCursors);
470	CursorInfo cursor1(vdp, vram, palette64_32768, 0x7fe08, 0x7ff80, cursorY, drawCursors);
471
472	if (cursor0.isVisible() \|\| cursor1.isVisible()) {
473	// raster background into a temporary buffer
474	std::array<uint16_t, 1024 + 3> buf; // allow to draw upto 3 pixels too many, e.g. see rasterBP2()
475	raster(colorMode, highRes,
476	IndexLookup(palette64_32768, palette256_32768),
477	vdp, vram,
478	subspan(buf, 0, dst.size()), x, y);
479
480	// draw sprites in this buffer
481	// TODO can be optimized
482	// TODO probably goes wrong when startX != 0
483	// TODO investigate dual palette in B4 and higher modes
484	// TODO check X-roll behavior
485	for (auto i : xrange(dst.size())) {
486	if (cursor0.dot()) {
487	if (cursor0.doXor) {
488	buf[i] ^= 0x7fff;
489	} else {
490	buf[i] = cursor0.color;
491	}
492	} else if (cursor1.dot()) {
493	if (cursor1.doXor) {
494	buf[i] ^= 0x7fff;
495	} else {
496	buf[i] = cursor1.color;
497	}
498	}
499	cursor0.shift();
500	cursor1.shift();
501	if ((cursor0.pattern == 0) && (cursor1.pattern == 0)) break;
502	}
503
504	// copy buffer to destination, translate from V9990 to host colors
505	for (auto i : xrange(dst.size())) {
506	dst[i] = palette32768[buf[i]];
507	}
508	} else {
509	// Optimization: no cursor(s) visible on this line, directly draw to destination
510	raster(colorMode, highRes,
511	PaletteLookup<Pixel>(palette64, palette256, palette32768),
512	vdp, vram, dst, x, y);
513	}
514	}
515
516	} // namespace openmsx

Keyboard